Cognitive Architectures for AI Agents

Why Agents Need Architecture

Modern AI agents are typically built by connecting an LLM to tools and giving it a system prompt that describes its role. This works for simple tasks, but breaks down as agents become more complex. Without a structured memory system, the agent forgets everything between sessions. Without goal management, it loses track of multi-step tasks. Without learning, it makes the same mistakes repeatedly. Without attention management, it wastes context window space on irrelevant information.

Cognitive architectures solve these problems by providing a principled framework for each capability. Instead of inventing ad-hoc solutions (store everything in a text file, summarize the conversation periodically, keep a to-do list in the prompt), the agent uses validated algorithms for memory storage and retrieval, attention allocation, goal decomposition, and learning from feedback. The architecture ensures that these components work together coherently rather than interfering with each other.

The oldest and most thoroughly studied cognitive architectures come from cognitive science: ACT-R, SOAR, and CLARION. These were designed to model human cognition, but their components map directly to the capabilities that AI agents need. Memory retrieval, attention, learning, and goal management are not uniquely human problems. They are general computational problems that these architectures have been solving for decades.

The Major Cognitive Architectures

ACT-R: Memory-Centered Architecture

ACT-R organizes cognition around declarative memory retrieval. The central cycle involves setting goals, retrieving relevant knowledge from declarative memory, and applying production rules to advance toward the goal. What makes ACT-R distinctive is its activation-based retrieval system, where each piece of knowledge has a dynamically computed accessibility score based on recency, frequency, and contextual associations. This means the agent automatically surfaces the most relevant knowledge for the current situation without needing explicit search queries.

For AI agents, ACT-R's memory model means the agent naturally develops expertise over time. Frequently used knowledge becomes more accessible. Recently updated information supersedes older versions. Contextually related knowledge surfaces together. The agent becomes more effective with use, not because it has been retrained, but because its memory dynamics naturally prioritize the most useful information.

SOAR: Problem-Solving Architecture

SOAR organizes cognition around search through problem spaces. The agent identifies the current state, proposes operators to transform it, selects the best operator, and applies it. When the agent encounters an impasse (no operator applies, or multiple operators compete), it creates a subgoal to resolve the impasse. Solutions to subgoals are compiled into new rules (chunks) that apply directly in future situations, enabling the agent to learn from experience.

For AI agents that need to handle complex, multi-step tasks, SOAR's subgoaling mechanism provides structured planning without requiring the agent to plan everything upfront. The agent can start working on a task and create subgoals as it encounters obstacles, learning from each resolution for future tasks.

CLARION: Dual-Process Architecture

CLARION models two levels of processing: a bottom level of implicit, subsymbolic processing (neural networks) and a top level of explicit, symbolic processing (rules). The bottom level learns through reinforcement and develops intuitive patterns. The top level extracts explicit rules from the bottom level's learned patterns. This dual-process design captures how experts develop intuitive judgment that goes beyond their articulated rules.

For AI agents, CLARION's framework suggests a design where the agent has both explicit knowledge (documented rules, procedures, facts) and implicit patterns (learned from usage data, interaction history, and outcome feedback). The explicit level provides interpretable, auditable behavior, while the implicit level captures nuances that are hard to express as rules.

Cognitive Architecture Concepts for Agent Design

Even if you do not adopt a full cognitive architecture, several concepts from the field improve agent design:

Working Memory vs Long-Term Memory

Cognitive architectures distinguish between working memory (the currently active information that the agent is operating on) and long-term memory (the full store of knowledge that the agent has accumulated). In AI agents, the context window serves as working memory and the external memory system serves as long-term memory. The architecture's job is to manage what moves between them: retrieving relevant long-term memories into working memory when needed and storing important working memory contents into long-term memory for future use.

Activation-Based Retrieval

Rather than retrieving everything that matches a query, activation-based retrieval ranks results by their dynamic accessibility. A memory's activation depends on how recently it was used, how often it has been useful, and how well it connects to the current context. This produces retrieval behavior that adapts to usage patterns without any explicit tuning, because the activation values update automatically with every interaction.

Goal Stack Management

Agents working on complex tasks need to manage multiple goals and subgoals without losing track of the overall objective. Cognitive architectures maintain a goal stack (or goal graph) that tracks the current goal, any subgoals created to handle obstacles, and the relationships between them. This prevents the common failure mode where an agent gets sidetracked by a subtask and forgets to return to the main task.

Production Rules and Procedural Knowledge

Knowledge about how to do things (procedural knowledge) is stored separately from knowledge about facts (declarative knowledge). Production rules encode condition-action pairs: if the current situation matches these conditions, take this action. This separation means the agent can acquire new facts without disrupting its skills, and can learn new skills without forgetting its facts.

Applying Cognitive Architecture to Modern AI Agents

A modern AI agent using cognitive architecture principles might look like this:

• LLM as the reasoning engine: The LLM handles language understanding, generation, and flexible reasoning. It fills the role of ACT-R's production matching system, selecting appropriate actions based on the current goal and retrieved context.
• External memory with ACT-R scoring: Long-term memory is stored in a vector database with additional metadata for activation computation. Retrieval uses cognitive scoring to rank results by recency, frequency, contextual relevance, and confidence.
• Context window as working memory: The LLM's context window holds the current goal, retrieved memories, tool results, and conversation history. Memory management ensures the most relevant information occupies the limited working memory space.
• Tool use as motor output: The agent's tools (code execution, file operations, API calls) serve as its interface with the external world, analogous to the motor module in cognitive architectures.

This design gives the agent structured memory management, contextually appropriate retrieval, and learning through usage, all without requiring modifications to the underlying LLM.

The Future of Cognitive Architectures in AI

As AI agents take on more complex, long-running tasks, cognitive architecture principles become increasingly important. An agent that assists a developer over months of a project needs to remember decisions, learn preferences, forget outdated context, and connect related knowledge across sessions. These are exactly the problems that cognitive architectures have studied for forty years.

The combination of modern LLMs (powerful reasoning and language capabilities) with cognitive architecture principles (structured memory, activation-based retrieval, learning through use) produces agents that are both flexible and grounded. The LLM provides adaptability and language understanding. The architecture provides the memory management and learning dynamics that make the agent effective over time rather than just in the current session.